Complete (or near-complete) non-archival research projects and exploratory work that I presented through posters or talks.
Examining the Structure of Text-Based Interventions in Large Language Models
Sweta Karlekar, Gabriel Chuang, and Claudia Shi · Columbia University Computer Science
Tests the linear representation hypothesis by treating concept changes in text as interventions and LLM activations as potential outcomes. Using WordNet antonyms and GPT-4-generated counterfactual sentence pairs, we compare constant shifts, linear transformations, and nonlinear mappings. Preliminary results favor constant shifts on a small sample of concepts.
Modeling the Projected Effect of Climate Change on the Prevalence of Food and Waterborne Illness Outbreaks
Nicolas Beltran and Sweta Karlekar · Columbia University Computer Science
Combines climate projections and probabilistic models within a causal-inference framework to explore food and waterborne illness outbreaks under different climate scenarios. The project compares outbreak models, examines uncertainty, and discusses limitations in data and causal assumptions.
Exploring a Flexible Computational Method for Comparing Massive Interaction Data from Science Visualizations
Sweta Karlekar · UNC Chapel Hill School of Education
Faculty advisor: Kelly Ryoo. Graduate student mentors: Emily Toutkoushian and Amanda Swearingen.
Developed a flexible method for comparing interaction logs from science visualizations using integer encoding and Levenshtein distance, identifying patterns across student knowledge groups to inform future personalized feedback.
A small selection of presentations that I’m particularly proud of.
A Mechanistic Interpretability Lens on Many Shot Jailbreaking
Presented for LLM Interpretability and Alignment. Explores hypotheses about many-shot jailbreaking through in-context learning, refusal directions, and persona shifts, connecting interpretability research with proposed experiments.
Graphical Models, Exponential Families, and Variational Inference — Chapters 1–2
Notes on Martin J. Wainwright and Michael I. Jordan (2008), covering graphical-model factorizations, message passing, junction trees, and approximate inference.
From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence
Notes on the 2026 paper by Marc Finzi, Shikai Qiu, Yiding Jiang, Pavel Izmailov, J. Zico Kolter, and Andrew Gordon Wilson. Topics include entropy, Kolmogorov complexity, time-bounded minimum description length, and epiplexity.